FLOPS in ResNet50_ImageNet

[miltonmondal]

Getting 4.12B flops using your code whereas almost all research papers mentioned 4.09B flops for this configuration

(pytorch default 76.15% test accuracy for pretrained model)

Can you please modify the code or mention the reason for getting 0.03B increase in FLOPs?

[jkhu29]

fvcore will get the output of 4.09G, but it will also print

Skipped operation aten::batch_norm 53 time(s)
Skipped operation aten::max_pool2d 1 time(s)
Skipped operation aten::add_ 16 time(s)
Skipped operation aten::adaptive_avg_pool2d 1 time(s)

Perhaps those papers ignore the computation of some of the operators.

[sovrasov]

@jkhu29 you're right! ptflops also considers batch norms and poolings as non-zero ops, that's why it outputs slightly greater numbers than expected.

[oyryrookie]

@jkhu29 you're right! ptflops also considers batch norms and poolings as non-zero ops, that's why it outputs slightly greater numbers than expected.

How to ignore the batch norms and poolings ops in ptflops?

[sovrasov]

@jkhu29 you're right! ptflops also considers batch norms and poolings as non-zero ops, that's why it outputs slightly greater numbers than expected.

How to ignore the batch norms and poolings ops in ptflops?

You can pass all the nn.Modules to exclude via ignore_modules parameter, see https://github.com/sovrasov/flops-counter.pytorch?tab=readme-ov-file#usage-tips-1